The 16x Context Trick: How Latent Context Models Finally Made Compression Work